X Chinese Topic Brief¶
Translation
English translation of the Chinese original.
Today's Top Picks¶
Top Pick 1 | NetEase Youdao's "欧拉讲题" (Euler Tutoring): a math-explaining Agent that draws on the board and takes follow-ups¶
Worth doing now: the product went live around November 5, so it was not released today; what is genuinely new is that Chinese-language X saw two independent hands-on tests of roughly 300 likes each on November 13, with angles aimed squarely at parents and non-professional tutors. 小互 (Xiaohu) @xiaohu original post (2026-11-13 15:13, 311 likes, 29.9K views, 46 reposts, 34 replies, 300 bookmarks) noted that it plans first, then turns the explanation into board writing, diagrams, animations and voice instructions; 铁锤人 (Tiechuiren) @lxfater original post (2026-11-13 11:28, 303 likes, 21.9K views, 56 reposts, 48 replies, 361 bookmarks) called it an AI Agent built specifically for math tutoring, with multiple agents working together to read the problem, plan and explain, and able to swap in different numbers and change conditions to generate variant problems to check whether the child really understands. Both stress that geometry auxiliary lines are drawn along with the explanation, and that the explanation can be interrupted with follow-up questions at any time.
Official information: the site 欧拉讲题 carries the title "For math problems you can't figure out, go to 欧拉讲题 (Euler Tutoring)" (queried 2026-11-14, HTTP 200); the 多知网 (Duozhi) report of 2026-11-05 says NetEase Youdao officially announced that this interactive math-explaining Agent is already available in app stores, with step-by-step explanation, interactive follow-ups, dynamic board writing and graphic demonstration; 京报网 (Beijing News) 2026-11-05 says the same.
Best angle to add: what parents really care about is not "can the AI solve the problem" but "will the child be able to solve it after the explanation." One can contrast question-search apps that hand over the answer directly with explaining while drawing, allowing follow-ups, and then setting variant problems. lxfater said "I took a look, this software is still free" — that is a user's statement; whether it is free is pending verification and must not be written as an official commitment.
Format suggestion: a standalone original post, or quote one of the hands-on tests and add a checklist for "how to tell whether the child has really learned it." No need to advertise on the vendor's behalf; just offer observations and questions.
Status: Ready to draft; the product's existence is Verified, whether it is still free is Pending verification; individual experience is not written up as a general review.
Top Pick 2 | DeepSeek turns App capabilities into an API and open-sources a chat front-end demo¶
Worth doing now: the official team open-sourced the core interactions of its web chat. Geek @geekbb original post (2026-11-13 22:32, 65 likes, 13.7K views, 9 reposts, 4 replies, 61 bookmarks) said there is no need to build a backend, the browser connects straight to the API, React + TypeScript, including streaming output, deep thinking, web search, attachments and conversation branching. 向阳乔木 (Xiangyang Qiaomu) @vista8 original post (22:44, 79 likes, 15.2K views, 9 reposts, 21 replies, 47 bookmarks) and 歸藏 (Guizang) @op7418 original post (22:26, 55 likes, 13.6K views, 1 repost, 15 replies, 26 bookmarks) were also relaying it, the latter quoting the DeepSeek official post.
Verified: the official news261113 note says App capabilities are now usable through the API; V4.1-Flash (deepseek-flash) supports web search, deep-thinking search, file input, link reading and safety checks on Messages, Responses and Chat Completions, with tool calls and context assembly handled server-side, and Search, Fetch, OCR, ASR and TTS also callable separately. The front-end demo has a demo, a code repository and a build guide. Public-facing applications must enable safety checks.
Best angle to add: for people who want to build their own assistant or knowledge base, this means no longer having to stitch search, link reading, voice and OCR together from scratch; but it must be made clear that tool calls are billed separately, not just tokens. According to Models & Pricing and the per-item guides (consulted 2026-11-14): Search costs USD 0.75 per 1,000 successful requests; Fetch costs USD 0.75 per 1,000 successful requests, with failures, empty results and timeouts not billed; OCR / file parsing costs USD 0.225 per 1,000 pages; TTS is listed at 0.45 per 10,000 characters with a limited-time 50% discount to USD 0.225 (Chinese characters count as 2 characters, other characters as 1); ASR is listed at USD 0.116 per hour of audio with a limited-time 50% discount to 0.058, billed by the second with any partial second counted as a full second, maximum 120 seconds and 8 MiB per file. The USD 0.058 per hour is the limited-time 50%-off price, not a long-term official price.
Format suggestion: a standalone original post, with a price list and the applicable boundaries.
Status: Ready to draft; the key prices are Verified.
Top Pick 3 | Follow-up: Chinese hands-on reports on Gemini 4 Hydrogen — some are swapping it in for Opus, others keep Carbon for writing¶
Worth doing now: release, pricing and first tests were written up on November 13; this round reports only the increments. Gorden Sun @Gorden_Sun original post (2026-11-13 10:48, 263 likes, 32.1K views, 16 reposts, 69 replies, 55 bookmarks) said Hydrogen "is actually Gemini 4 Flash, but with a notable capability jump," that its text carries on from 3.8 Flash with less of an AI flavor; in his "大肥鱼" (Big Fat Fish) video workflow it can directly replace Opus 5.5, saving roughly half when tokens are converted into dollars; his own task contexts run about 300K, and he is unsure whether results degrade on longer tasks; it is usable with a Gemini Pro subscription. 向阳乔木 (Xiangyang Qiaomu) @vista8 original post (2026-11-13 23:03, 75 likes, 11.1K views, 3 reposts, 23 replies, 59 bookmarks) relayed a small-group discussion, saying Gemini 3.5 flash-lite is the most cost-effective multilingual translation model, that Gemini 4 Carbon is the best Chinese writing model but a bit slow, and noting that this is a claim he is about to go and verify.
Best angle to add: use the Verified pricing to help readers divide the work. Hydrogen is USD 1.50 per million input tokens and 7.50 output; Carbon's early-access price is 2 input and 10 output, becoming 4 / 20 after November 30. Use the fast tier for batch work, translation and running pipelines, and keep Carbon for small-volume, high-value tasks like Chinese writing.
Format suggestion: a standalone original post. Status: Ready to draft; pricing is Verified, while the experience is a personal workflow self-test and small-group claims, not a review conclusion.
Candidate Topics¶
1. Two Chinese hands-on tests of NetEase Youdao's "欧拉讲题"¶
Original posts and engagement: 小互 (Xiaohu) @xiaohu original post, 2026-11-13 15:13, 311 likes, 29.9K views, 46 reposts, 34 replies, 300 bookmarks; 铁锤人 (Tiechuiren) @lxfater original post, 2026-11-13 11:28, 303 likes, 21.9K views, 56 reposts, 48 replies, 361 bookmarks.
Content and focus: both describe geometry auxiliary lines drawn along with the explanation, interruptions with follow-up questions, and generated variant problems; xiaohu stresses planning first and then generating board writing and voice instructions, lxfater stresses multi-agent coordination and the self-test method of "swapping numbers and changing conditions."
Why it's worth doing / possible angles: this item corresponds to "Today's Top Picks 1." It can be written from the parent's perspective as "will the child be able to do it after the explanation," rather than comparing who computes faster.
Timeliness and verification: within 24 hours. The product has been live since November 5; the official site and two media reports are Verified; the free-of-charge claim comes from lxfater and is Pending verification.
2. DeepSeek App capabilities exposed as an API, plus an open-source chat front-end¶
Original posts and engagement: 向阳乔木 (Xiangyang Qiaomu) @vista8 original post, 79 likes, 15.2K views, 47 bookmarks; Geek @geekbb original post, 65 likes, 13.7K views, 61 bookmarks; 歸藏 (Guizang) @op7418 original post, 55 likes, 13.6K views, 26 bookmarks. All between 2026-11-13 22:26 and 22:44.
Content and focus: all three put the emphasis on "the front-end runs directly, no need to build your own backend"; geekbb listed five capabilities: streaming, deep thinking, web search, attachments and conversation branching.
Why it's worth doing / possible angles: this item corresponds to "Today's Top Picks 2." The added value is the itemized price list and the easily confused point that "tool fees are on top of token fees."
Timeliness and verification: within 24 hours. The official note, repository and pricing page are Verified; the TTS and ASR discounts are limited-time prices.
3. Follow-up: swapping Hydrogen in for Opus, keeping Carbon for writing¶
Original posts and engagement: Gorden Sun @Gorden_Sun original post, 2026-11-13 10:48, 263 likes, 32.1K views, 16 reposts, 69 replies, 55 bookmarks; 向阳乔木 (Xiangyang Qiaomu) @vista8 original post, 2026-11-13 23:03, 75 likes, 11.1K views, 3 reposts, 23 replies, 59 bookmarks.
Content and focus: one is a substitution-and-savings self-test in a video workflow, the other is a small-group claim about dividing work between models; both point to "picking the tier by task" rather than "who is strongest."
Why it's worth doing / possible angles: this item corresponds to "Today's Top Picks 3." Use the pricing table to convert personal impressions like "saves half" into checkable terms.
Timeliness and verification: within 24 hours; a Follow-up, since the release and first tests were already written up on November 13. Pricing is Verified; replacing Opus and the small-group ranking are personal claims, not written as conclusions.
4. Claude Managed Agents can be self-hosted, but the models do not move by default¶
Original posts and engagement: Baye (贝叶) @waylybaye original post, 2026-11-13 11:52, 88 likes, 30.5K views, 3 reposts, 20 replies, 81 bookmarks, quoting the ClaudeDevs explanatory post. There are also some low-engagement explanatory posts in the Chinese-language community, which need not all be listed.
Content and focus: readers are most easily carried away by the words "self-hosted," assuming the Claude models themselves can be moved wholesale into their own data center.
Why it's worth doing / possible angles: separate the three layers — self-hosting only the sandbox; self-hosting the service as well; and the model still running at Anthropic or at a provider you choose. According to the official documentation, the service and sandbox can be placed in your own cloud or data center, the model runs at Anthropic or the chosen provider, downloading and use require applying through the Anthropic account team, and Amazon Bedrock, Claude on Google Cloud, Microsoft Foundry or a self-built gateway are options; this is not the same tier as hosting only the sandbox (with the service still run by Anthropic), and the differences are set out in Differences from the managed service. The official pages do not publish individual pricing or an official general-availability date.
Timeliness and verification: within 24 hours. The documentation is Verified; specific onboarding conditions are subject to the Anthropic account team's reply.
5. DeepSeek Harness wants Agents to pick up work themselves: dsh-superagent is a research prototype¶
Original posts and engagement: 小墨同学 (Xiao Mo) @xiaomovps original post, 2026-11-13 09:54, 91 likes, 11.1K views, 10 reposts, 14 replies, 91 bookmarks; GoGoFly23 @GoGoFly23 original post, 2026-11-12 21:52, 168 likes, 6326 views, 19 reposts, 11 replies, 115 bookmarks.
Content and focus: xiaomovps describes integration with GitHub and Feishu, a Router + Worker structure, minimal interruption, and delivery via PR, and states plainly that this is still a research prototype, mainly around a single GitHub repository, and requires self-deployment. GoGoFly23's is a personal demo: using DeepSeek Harness over MCP to take over SolidWorks, with natural-language modeling and assembly — this is the author's own account, not an official feature release.
Why it's worth doing / possible angles: the way of working shifts from "open a window and give instructions" to "@ it in an Issue or a group chat, and the result comes back as a PR." At the same time it must be stressed that the repository itself says it is "not a formal product."
Timeliness and verification: one item within 24 hours, one at 24–48 hours. The GitHub API, queried 2026-11-14 08:29 (Shanghai time), shows the repository at 23,018 stars under the MIT license, so one can write "about 23K stars (queried 2026-11-14)"; the README states it is a research prototype for reference only, not an official DeepSeek product, and not suitable for direct use as production software, that the code grew out of the team's day-to-day development with a large amount of the engineering done by coding agents, and that it was adapted before release and is not entirely identical to the internal version or the paper's description.
6. Follow-up: the Chinese-language spread of Edge0-145B — don't mix up the numbers¶
Original posts and engagement: @QwenAIFans original post, 2026-11-13 12:39, 272 likes, 19.5K views, 20 reposts, 10 replies, 237 bookmarks.
Content and focus: the post says 145 billion parameters, that the official team measured 121 tokens per second on a 24GB 5090D, that the base is Qwen3.8 Flash-Next with roughly 10 billion parameters activated per word, and it links a model description.
Why it's worth doing / possible angles: the daily brief already covered the release and the vendor's self-test, so here only the Chinese-language discussion angle is added, along with a correction of one easily confused figure. According to the official blog, 145B and 120B were released on November 12, peak memory is about 29GB, and the 1100+ tokens/s prefill and 80+ tokens/s decode are both vendor self-tests; the license is the Edge0 Community Model License 1.0, not Apache 2.0, and the "Commercial Model Offering" requires separate authorization.
Timeliness and verification: within 24 hours; a Follow-up. "24GB 5090D, 121 tokens/s" and the blog's "29GB, 80+ decode" do not use the same basis, so mark this Pending verification and do not write it as a confirmed official figure; self-reported benchmarks are not written as independent reproduction.
7. Worth watching: using Pi's Virtual Models to route domestic models automatically¶
Original posts and engagement: 小墨同学 (Xiao Mo) @xiaomovps original post, 2026-11-13 17:09, 127 likes, 7871 views, 11 reposts, 12 replies, 105 bookmarks.
Content and focus: relays a setup from Hugging Face co-founder Julien Chaumond — using Pi's Virtual Models / Extension to combine a local Qwen 3.8 27B with remote Kimi K3 and GLM-5.3 into hybrid / auto virtual models, routing automatically by task complexity or local machine load, without changing Pi's core code. He himself also often switches manually between GPT and DeepSeek, and considers the approach worth a look, though automatic switching is not necessarily cheaper.
Why it's worth doing / possible angles: suits a piece on the "local small model + remote large model" routing idea, making clear that whether what is saved is waiting time or money has to be looked at separately.
Timeliness and verification: within 24 hours. The discussion sample is thin, so mark it Worth watching; Julien's original post was not verified in this round, so write it as "a post relays"; the savings effect is Pending hands-on test.
8. Follow-up: Stanford CS329Z course videos, 4 lectures now public¶
Original posts and engagement: meng shao @shao__meng original post, 2026-11-13 09:57, 188 likes, 6262 views, 32 reposts, 7 replies, 130 bookmarks.
Content and focus: the lecture notes have long been public; the increment this round is that videos are starting to appear, and the post says 4 lectures are already available.
Why it's worth doing / possible angles: course materials going public was already reported on November 12, so here only the increments: the order to follow, each lecture's topic, and what background each suits.
Timeliness and verification: within 24 hours; a Follow-up. The course homepage is still cs329z.stanford.edu, and "4 lectures already public" follows that post, and if the video list cannot be opened and counted item by item, it should be marked as the post's claim with the specific lecture count Pending verification. Do not copy the November 12 engagement figures again as this round's heat.
This issue selects 8 items, no padding.
Coverage gap: some accounts on the watch list could not be confirmed as having posted this round, and 3 of the sweeps paged through only one page because of posting volume; this does not affect the factual judgments on the topics collected above.